0

Is it usefull to check and include the lags a variable would use for an univariate AR Model into an multivariate RNN? Or is this decked through the 3D sturcture of my input data?

Tim
  • 108,699
  • 20
  • 212
  • 390
N.Messers
  • 23
  • 6
  • Try it. If it works, you would end up with a catchy name for a novel network for a publication: ARRNN ‍☠️ – Tim Jun 25 '21 at 11:36
  • Okay, this got to be the funniest answer I've ever seen on this webside, kudos! :D – N.Messers Jul 01 '21 at 11:10

1 Answers1

0

I am not aware of such a model, but you should try it. Quite a similar model, a hybrid of exponential smoothing and RNN won the M4 forecasting competition (Makridakis et al, 2018). Exponential smoothing same as ARIMA, is another family of classical time-series forecasting models. If it worked, there is no reason not to try combining the AR model with RNN.

If you think about it, adding AR component to RNN is a quite similar idea to what ResNet (He et al, 2015) did by passing the information from previous layers to the subsequent ones.

enter image description here

AR model takes information from past time points to make the prediction. RNN takes the information from the current input and the output of the previous recursive step. If you combine them (e.g. concatenate and pass them together to the next layer) you will combine the transformed and untransformed information, similar as ResNet did.

Tim
  • 108,699
  • 20
  • 212
  • 390